Results - Details
Search command : Author="Τσακαλίδης"
And Author="Παναγιώτης"
Current Record: 14 of 68
|
Identifier |
000441260 |
Title |
Exploitation of noisy automatic data annotation for CNN training and its application to hand posture classification |
Alternative Title |
Εκμετάλλευση αυτόματης επισημείωσης δεδομένων για εκπαίδευση Συνελικτικών Νευρωνικών Δικτύων και εφαρμογή στην κατηγοριοποίηση χειρομορφών |
Author
|
Λυδάκης, Γεώργιος Ε.
|
Thesis advisor
|
Αργυρός, Αντώνιος
|
Reviewer
|
Τσακαλίδης, Παναγιώτης
Ρούσσος, Αναστάσιος
|
Abstract |
In recent years, advances in deep learning have resulted in a large-scale revolution in the
field of Artificial Intelligence. Deep learning methods have been successfully applied to a
variety of research topics, ranging from natural language processing and bioinformatics
to speech recognition and computer vision, with the common goal of inferring a function
which maps an input domain to a target one. However, the success of deep models at
inferring such a function usually relies on the existence of a large amount of annotated
training data, that is, input samples for which the output is specified. Due to the
requirement for large such training datasets, significant research is being conducted on
methods for reducing the human effort that is necessary for their annotation.
Semi-supervised approaches, methods for synthetic data generation, and techniques for
generating and handling automatic annotation are receiving increasing attention. In this
work, we investigate a technique for utilizing automatically annotated data in
classification problems. Using a small number of manually annotated samples, and a large
set of data that feature automatically created, noisy labels, our approach trains a
Convolutional Neural Network (CNN) in an iterative manner. The automatic annotations
are combined with the predictions of the network in order to gradually expand the
training set. This expansion attempts to select automatically annotated samples for which
the label is deemed to be correct.
The proposed approach is generic and can be applied to any classification problem. In
order to evaluate its performance, we apply it to the problem of hand posture recognition
from RGB images. In general, the observation and interpretation of the human hand is
highly useful in several application areas, so significant research has been carried out in
the topics of 3-D hand tracking, observation of interaction of hands with objects as well
as hand posture and gesture recognition. Sign Language Recognition is one area where
hand posture recognition is especially useful, as the postures of a signer's hands are
essential features in the translation of a sign language.
Motivated by the usefulness and impact of Sign Language Recognition, we develop a
method for automatically annotating images or videos of hand postures, and apply it to
the problem of classifying 19 postures that are common in the Greek Sign Language. The
manual annotation of such data is a time-consuming process. Their automatic annotation
is based on associating 3D joint configurations of hands with classes (hand posture labels),
and yields noisy such labels which can be used to train a Convolutional Neural Network.
These characteristics of the problem make it a suitable candidate for applying our
proposed method for automatic expansion of training datasets. We compare the results
of training a CNN classifier with and without the use of our technique. Our method yields
a significant increase in average classification accuracy, and also decreases the deviation
in class accuracies, thus indicating the validity and the usefulness of the proposed
approach.
|
Language |
English |
Subject |
Artificial neural networks |
|
Computer vision |
|
Deep learning |
|
Βαθιά μάθηση |
|
Τεχνητά νευρωτικά δίκτυα |
|
Υπολογιστική όραση |
Issue date |
2021-07-30 |
Collection
|
School/Department--School of Sciences and Engineering--Department of Computer Science--Post-graduate theses
|
|
Type of Work--Post-graduate theses
|
Permanent Link |
https://elocus.lib.uoc.gr//dlib/d/a/e/metadata-dlib-1626425689-23029-28441.tkl
|
Views |
629 |